MAE train & MAE test loss functions give same values over multiple epochs, effectively no training being done.

[ethan530]

Most of the code is from the pytorch_workflow.ipynb file, I changed the dataset to be google stock data I found online.
Google.csv

In the zip file is the code I have written.
v2.zip

If you run through the cells and get to the loss calculation, the test and training loss is:

Even if I change the number of epochs to a higher number, the calculations do not change, they remain at the same number instead of decreasing.

In graph form, it looks like this:

Does anyone have an idea of why this happens?

[pusapatiakhilraju]

The architecture of neural networks in "Workflow.ipynb" does not allow them to learn complex patterns in stock data. In the workflow notebook the neural network architecture consists only of linear layers and do not have activation functions in between the layers to account for non linearity. Hence, increasing the epochs is only one of the solution to improve the performance of the model . There are also many other ways to teach the model different patterns and in the future workbook Daniel specifies what other things could be changed in the architecture to make the model better. Hope this clarifies your question.

[mrdbourke]

This is correct!

One neural network doesn't necessarily fit all kinds of data.

Especially with stock data (which I wouldn't actually recommend using almost any kind of model to predict, as stocks are mostly unpredictable).

[ethan530]

Thanks for replying, the Google data when I used matplotlib to visualise the data looked like this:

Maybe its just because there are more than 4000 values on the graph, but the training and test data looks surprisingly linear.